A Gradient Bound for Free Boundary Graphs
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
De Silva, Daniela, and David Jerison. “A Gradient Bound for Free
Boundary Graphs.” Communications on Pure and Applied
Mathematics 64.4 (2011): 538–555. Web. 26 June 2012. © 2010
Wiley Periodicals, Inc.
As Published
http://dx.doi.org/10.1002/cpa.20354
Publisher
Wiley Blackwell (John Wiley & Sons)
Version
Author's final manuscript
Accessed
Thu May 26 00:30:04 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/71210
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike 3.0
Detailed Terms
http://creativecommons.org/licenses/by-nc-sa/3.0/
arXiv:1009.4694v1 [math.AP] 23 Sep 2010
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
D. DE SILVA AND D. JERISON
Abstract. We prove an analogue for a one-phase free boundary problem of
the classical gradient bound for solutions to the minimal surface equation. It
follows, in particular, that every energy-minimizing free boundary that is a
graph is also smooth. The method we use also leads to a new proof of the
classical mimimal surface gradient bound.
1. Introduction
n
Let Ω be a domain in R , and consider the one-phase free boundary problem
(u ≥ 0)
∆u = 0,
in Ω+ (u) := {x ∈ Ω : u(x) > 0},
(1.1)
|∇u| = 1, on F (u) := ∂Ω ∩ Ω+ (u).
The set F (u) is known as the free boundary. There are strong parallels between the
theory of these hypersurfaces and the theory of minimal surfaces. The existence of
solutions u and partial regularity (smoothness almost everywhere with respect to
surface measure) of the free boundary F (u) was proved by Alt and Caffarelli [AC].
We will begin by formulating our main result in the special case of energyminimizing solutions. We call u energy-minimizing on Ω̄ if u minimizes the functional
ˆ
J(v) = (|∇v|2 + χ{v>0} )dx,
Ω
among all functions with the same boundary values as u. The first variation (EulerLagrange) equations for u are (1.1). Indeed, it is easy to show that u is harmonic
in Ω+ (u), and it follows from deeper results of [AC, C1, C2] that u satisfies the free
boundary condition |∇u| = 1 on F (u) in a viscosity sense, defined in Section 2.
Define a cylinder of height 2L, with base the ball of radius Br in Rn−1 , by
C(r, L) := Br × (−L, L) ⊂ Rn ;
(and CL := C(1, L)).
Theorem 1.1. If an energy-minimizing solution u on the cylinder CL is monotone
in the vertical direction,
∂u/∂xn ≥ 0 on CL+ (u),
and its free boundary F (u) is a fixed distance from the top and bottom of the cylinder, i. e.,
F (u) ⊂ CL−ǫ for some ǫ > 0,
then F (u) is the graph of a smooth function ϕ,
F (u) = {(x, y) : x ∈ B1 ;
y = ϕ(x)}
Both authors were partially supported by NSF grant DMS-0244991.
1
2
D. DE SILVA AND D. JERISON
with
sup |∇ϕ(x)| ≤ C
x∈B1/2
for a constant C depending only on L, ǫ, and n.
Let us compare our theorem with the classical gradient bound on minimal surfaces due to Bombieri, De Giorgi, and Miranda, which can be stated as follows.
Theorem 1.2. [BDM] Let φ ∈ C ∞ (B1 ) be a solution to the minimal surface equation
!
∇φ
(1.2)
div p
= 0 in B1 ,
1 + |∇φ|2
with |φ| ≤ M. Then
(1.3)
|∇φ| ≤ C
in
B1/2
with C depending on n and M .
The hypothesis ∂u/∂xn ≥ 0 in Theorem 1.1 implies, by the strong maximum priniciple, that ∂u/∂xn > 0. Therefore the level surfaces {x : u(x) = c} for c > 0 are
graphs. The hypothesis that the free boundary is a fixed distance from the top and
bottom of the cylinder replaces the hypothesis in Theorem 1.2 that the oscillation of
the function φ is bounded by M . Furthermore, the minimal surface equation (1.2)
implies that the graph of φ is area-minimizing, so that the assumption in Theorem
1.1 that the free boundary is energy minimizing is analogous.
In the theory of minimal surfaces, it is well-known that minimal graphs are real
analytic in the interior of the their domain of definition. The key first step in
the proof of full regularity of the minimal graphs is to establish that the graph is
Lipschitz, that is, the graph of a function with a bounded gradient. The gradient
bound proved here leads, likewise, to full regularity. If the free boundary is a
Lipschitz graph, then Caffarelli [C1] proved that the graph is C 1,α for some α > 0.
Higher regularity results of [KN] then yield the local analyticity of F (u). So real
analyticity follows if one can confirm the Lipschitz property, i. e., the gradient
bound.
In [D2], an a priori gradient bound for smooth free boundary graphs is proved in
the case when n = 2, 3. The proof given there is also motivated by the strong analogy
with minimal surfaces, but is completely different. An advantage of the results here
is that because they work in all dimensions, they can be expected to apply to the
free boundary analogue of the Bernstein problem. The application we have in mind
is to the construction (as yet unrealized) of a global solution to the free boundary
problem (other than the obvious solution u(x) = x+
1 ) whose level surfaces are
graphs. This would be analogous to the counterexample to the Bernstein conjecture
— a complete non-planar mimimal graph constructed in [BDG] in R9 . In [DJ], it is
shown that a certain cone in R7 is the free boundary analogue of the Simons cone
in minimal surface theory. Based on this example, one should expect to find a free
boundary whose level surfaces are non-flat graphs in R8 .
The theorem whose proof occupies most of this paper has a more technical statement. See Section 2 for the definition of a viscosity solution and nontangentially
accessible (NTA) domains.
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
3
Theorem 1.3. Let u be a viscosity solution to (1.1) in the cylinder CL . Suppose
that u is monotone in the vertical direction,
∂u/∂xn ≥ 0 on CL+ (u),
and its free boundary is given as the graph of a continuous function ϕ, F (u) =
{(x, y) : x ∈ B1 ; y = ϕ(x)}. Suppose that the oscillation of ϕ is bounded,
max |ϕ(x)| ≤ L − 1,
x∈B1
and, finally, that there is a nontangentially accessible (NTA) domain D such that
1
9
∩ CL+ (u) ⊂ D ⊂ CL+ (u).
,L −
C
10
2
Then
sup |∇ϕ(x)| ≤ C
x∈B1/2
for a constant C depending only on L, the NTA constants, and n.
Theorem 1.1 will follow from Theorem 1.3 using results of [D1]. Roughly speaking, [D1] shows that the hypotheses of Theorem 1.1 imply the hypotheses of Theorem 1.3. In particular, a key estimate from [D1] is that the positive phase satisfies
an NTA property on any smaller cylinder. Moreover, it is also proved in [D1] that
under the hypotheses of Theorem 1.1, the free boundary is the graph of a continuous
function ϕ.
The proof of Theorem 1.3 is based on comparing u(x) to its vertical translates
u(x + ten ). One constructs a family of supersolutions related to u(x + ten ) and uses
a deformation maximum principle argument to show that u(x + ten ) ≥ u(x) + ct
for sufficiently small t > 0. The function u(x) is comparable to the distance from
x to the free boundary. The estimate shows that the change in u in the vertical
direction is comparable to the change in u in the direction normal to each level
surface, which is equivalent to a Lipschitz bound on the graph of the level surface.
The construction of the family of supersolutions makes use of the basic estimates
on NTA domains which were the reason the notion of NTA was introduced in [JK].
The NTA property guarantees that every positive harmonic function that vanishes
on the boundary vanishes at the same rate as u. The NTA property was first used
in connection with regularity of free boundaries by Aguilera, Caffarelli and Spruck
[ACS], who proved a partial regularity result. The NTA property also holds for the
singular conic solution of Alt and Caffarelli. (This cone is not a graph, of course.
Otherwise it would contradict Theorem 1.3.)
Our proof of the gradient bound for free boundaries leads to a new proof of the
classical gradient bound for minimal graphs. This new proof of Theorem 1.2 is
related to a much simpler proof due to N. Korevaar [K]. The hope is that this
new method, while more complicated than the method in [K], will ultimately apply
to classes of semilinear problems that include both free boundary problems and
minimal surface problems as singular limits. An interesting aspect of our proof is
that it deepens the analogy between minimal surfaces and free boundaries.
The paper is organized as follows. In Section 2 after briefly recalling some standard definitions and known results, we prove Theorem 1.3 and deduce Theorem 1.1.
We present our proof of Theorem 1.2 in Section 3. In Section 4, we examine the
4
D. DE SILVA AND D. JERISON
parallels between the two proofs and especially between two key parallel ingredients, namely the boundary Harnack inequality for NTA domains and the intrinsic
Harnack inequality of Bombieri and Giusti [BG].
2. Gradient bound for free boundary graphs
2.1. Preliminaries. We recall the definition of a viscosity solution [C1].
Definition 2.1. Let u be a nonnegative continuous function in Ω. We say that u is
a viscosity solution to (1.1) in Ω if and only if the following conditions are satisfied:
(i) ∆u = 0 in Ω+ (u);
(ii) If x0 ∈ F (u) and F (u) has at x0 a tangent ball Bǫ from either the positive
or the zero side, then, for ν the unit radial direction of ∂Bǫ at x0 into
Ω+ (u),
u(x) = hx − x0 , νi+ + o(|x − x0 |), as x → x0 .
Standard elliptic regularity theory implies that if F (u) is a smooth surface near
x0 , then u is smooth up to the free boundary near x0 and the free boundary
condition |∇u| = 1 is valid in the classical sense in such a neighborhood.
Denote by d(x) = dist(x, F (u)). In this section, the balls Br = Br (0) and Br (x)
will be in Rn while the balls Br (x) will be in Rn−1 . The following result follows
easily from the Hopf lemma and interior regularity of elliptic equations (see for
example [CS],[D2]).
Lemma 2.2. Let u be a viscosity solution to (1.1) in B1 , 0 ∈ F (u). Then, u is
Lipschitz continuous in B1/2 and there is a dimensional constant K such that
sup |∇u| ≤ K,
B1/2
and
u(x) ≤ Kd(x),
for all x ∈ B1/2 .
Definition 2.3. We say that a viscosity solution u is nondegenerate in B1 if there
is a constant c > 0 such that u(x) ≥ cd(x) for all x ∈ B1+ (u).
We now recall the notion of nontangentially accessible (NTA) domains.
Definition 2.4. A bounded domain D in Rn is called NTA, when there exist
constants M and r0 > 0 such that:
(i) Corkscrew condition. For any x ∈ ∂D, r < r0 , there exists y = yr (x) ∈ D
such that M −1 r < |y − x| < r and dist(y, ∂D) > M −1 r;
(ii) The Lebesgue density of Dc at any of its points is bounded below uniformly
by a positive constant c, i.e for all x ∈ ∂D, 0 < r < r0 ,
|Br (x) \ D|
≥ c;
|Br (x)|
(iii) Harnack chain condition. If ǫ > 0 and x1 , x2 belong to D, dist(xj , ∂D) > ǫ
and |x1 − x2 | < C1 ǫ, then there exists a sequence of C2 balls of radius
cǫ such that the first ball is centered at x1 , the last at x2 , such that the
centers of consecutive balls are at most cǫ/2 apart. The number of balls
C2 in the chain depends on C1 , but not on ǫ.
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
5
We recall some results about NTA domains [JK]. We start with the following
boundary Harnack principle for harmonic functions.
Theorem 2.5. (Boundary Harnack principle) Let D be an NTA domain and let
V be an open set. For any compact set K ⊂ V, there exists a constant C such that
for all positive harmonic functions u and v in D vanishing continuously on ∂D ∩ V,
and x0 ∈ D ∩ K,
C −1
v(x0 )
v(x0 )
u(x) ≤ v(x) ≤ C
u(x), for all x ∈ K ∩ D.
u(x0 )
u(x0 )
The boundary Harnack inequality above will be our main tool in the proof of
Theorem 1.3. We will also need some further facts. First, recall that for any
bounded domain D ⊂ Rn and any arbitrary y0 ∈ D, one can define the harmonic
measure ω y0 of D evaluated at y0 (for the definition see for example [JK]). We
note that for any y1 , y2 ∈ D, the measures ω y1 and ω y2 are mutually absolutely
continuous. Hence, from now on we fix a point y0 ∈ D and denote ω = ω y0 .
A nontangential region at x0 ∈ ∂D is defined as
Γα (x0 ) = {x ∈ D : |x − x0 | < (1 + α)dist(x, ∂D)}.
Let u be defined on D and f on ∂D. We say that u converges to f nontangentially
at x0 ∈ ∂D if for any α,
lim u(x) = f (x0 ) for x ∈ Γα (x0 ).
x→x0
The following Fatou-type theorem was proved in [JK].
Theorem 2.6. Let D be an NTA domain. If u is a positive harmonic function in
D, then u has finite nontangential limits for ω-almost every x0 ∈ ∂D.
We deduce from this the following regularity result for NTA free boundaries.
Lemma 2.7. Let u be a viscosity solution to (1.1) in B1 , u non-degenerate in B3/4 ,
and 0 ∈ F (u). Assume that there is an NTA domain D such that D ⊂ B1+ (u) and
F (u) ∩ B3/4 ⊂ ∂D. Then, F (u) ∩ B1/2 is smooth almost everywhere with respect to
harmonic measure ω of D.
Proof. Since each partial derivative ∂u/∂xj is a bounded harmonic function, Theorem 2.6 implies that for ω-almost every x0 ∈ F (u) ∩ B1/2 , there exists a ∈ Rn
such that for every α < ∞, ∇u(x) → a as x → x0 , for x ∈ B1+ (u), |x − x0 | <
(1 + α)dist(x, F (u)). We will prove that F (u) is flat and hence smooth in a neighborhood of x0 . The idea of the proof is to show that for x near x0 , u is close
to a linear function with gradient a. Provided that a is not the zero vector, this
will show us that the level sets of u are flat and hence (by [AC, C2]) that the free
boundary is smooth near x0 .
For notational simplicity assume x0 = 0. Denote by ur the rescaling of u,
ur (x) = u(rx)/r. We will use the notation A1 ≈ A2 for positive numbers that are
comparable modulo constants that depend only on the NTA constants and the ratio
of u(x) to the distance to the free boundary (bounded above and below by Lemma
2.2 and nondegeneracy). Consider a point z ∈ B1+ (ur ) such that ur (z) ≈ 1. Note
that although the point z depends on r, we require the constants in comparability
of ur (z) with 1 to be independent of r as r → 0. For any x ∈ B1 ∩ {ur > 0},
6
D. DE SILVA AND D. JERISON
the NTA properties imply there is a (nontangential, corkscrew) path p(t) such that
p(0) = x, p(1) = z, |p′ (t)| ≤ C and
ur (p(t)) ≈ dist(p(t), F (ur )) ≈ t + ur (p(0))
independent of r.
Fix C2 << C1 < ∞, δ > 0, and denote,
+
(ur ) : ur (x) ≥ δ},
Tδr = {x ∈ BC
1
+
Γrα = {x ∈ B1/r
(ur ) : |x| < (1 + α)dist(x, F (ur ))}.
Since ur (x) is comparable to the distance from x to F (ur ), for any x ∈ Tδr ∩ BC2 ,
r
there is a constant c > 0 such that the path p(t) from x to z belongs to Tcδ
. Choose
α sufficiently large depending on δ and c > 0 and r sufficiently small depending on
C1 such that
r
Tcδ
⊂ Γrα .
Thus there is r0 > 0 (depending on C1 , δ, and α) such that that for r < r0 ,
|∇ur (x) − a| < δ,
r
.
for x ∈ Tcδ
Define a linear function of x, by L(x) = ur (z) + a · (x − z). For all x ∈ Tδr ∩ BC2 ,
r
since ur (z) − L(z) = 0, and p(t) ∈ Tcδ
,
ˆ 1
|ur (x) − L(x)| = (∇ur (p(t)) − a) · p′ (t)dt ≤ C3 δ.
0
In all, we have shown that for every x ∈ BC2 such that ur (x) ≥ δ,
|ur (x) − L(x)| ≤ C3 δ.
Next, we deduce that |a| ≈ 1. (The upper bound |a| ≤ K already follows from
the upper bound on |∇u|.) Since ur (0) = 0 and ur (z) ≈ 1, for some 0 < t < 1, the
point x = tz satisfies ur (x) = δ. So x ∈ Tδr ∩ BC2 and |δ − ur (z) − a · (x − z)| ≤ C3 δ.
Hence, |a| ≥ |a · (x − z)| ≥ ur (z) − δ − C3 δ ≥ ur (z)/2. (All we need in what follows
is that a is bounded and nonzero.)
We can now conclude that the free boundary is flat in the appropriate sense.
Consider a point x ∈ F (ur ) ∩ B1 and its path p(t) to z. There is t > 0 such that
ur (p(t)) = δ. Denote y = p(t). Then |y − x| ≤ Cδ and y ∈ Tδ ∩ BC2 . The preceding
argument says |ur (y) − L(y)| ≤ C3 δ. Therefore,
|L(x)| ≤ |L(y)| + |L(x) − L(y)| ≤ |ur (y) − L(y)| + |ur (y)| + |a · (x − y)| ≤ C4 δ
for a larger constant C4 . Since a is bounded away from 0 in length, the bound on
L(x) implies that every point of F (ur ) ∩ B1 is within a distance a constant times
δ of the plane L(x) = 0. For sufficiently small δ, this flatness condition implies
smoothness of the free boundary (see [AC, C2]).
2.2. The proof of Theorem 1.3. Throughout the proof, ci , Ci denote constants
depending on L, n, and possibly on the NTA constants. Also, a point x ∈ Rn may
be denoted by (x′ , xn ), with x′ = (x1 , . . . , xn−1 ).
We divide the proof in three steps.
Step 1: Nondegeneracy and separation of level sets.
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
7
We show first the nondegeneracy of u, namely that if Bρ (x0 ) ⊂ CL+ (u), ρ < 1,
then
(2.1)
u(x0 ) ≥ γn ρ
for a dimensional constant γn > 0.
Denote by g a strictly superharmonic function on the annulus E = B2 \B1 such
that


on ∂B2 ,
 g = an
g=0
on ∂B1 ,


|∇g| < 1 on ∂B1 ,
with an > 0 small dimensional constant. Let r = ρ/4. Denote gr (x) = rg(x/r),
and
ht (x) = gr (x − x0 − ten )
defined on the closed annulus Et = B̄2r (x0 + ten )\Br (x0 + ten ). For t sufficiently
small, Et ⊂ {x : −L < xn < −L + 1} so that ht (x) ≥ 0 = u(x) for x ∈ Et .
Increasing t translates the region Et upwards. Let t0 be the least t for which
the graph of ht touches the graph of u, i. e., so that there is a point z0 ∈ Et
for which ht (z0 ) = u(z0 ) > 0. Because ht is a strict supersolution the point z0
belongs to the outer boundary, z0 ∈ ∂B2r (x0 + t0 en ). Furthermore, because the
free boundary of u and ht can’t touch, t0 ≤ −ρ − r < 0. Monotonicity of u implies
u(z0 − t0 en ) ≥ u(z0 ) = ht0 (z0 ) = an r. Finally, since |z0 − t0 en − x0 | = 2r = ρ/2,
Harnack’s inequality comparing the value of u at z0 −t0 en and x0 implies that there
is a dimensional constant γn > 0 such that u(x0 ) ≥ γn ρ, as required.
Next, we will show that level sets near the top of the cylinder are separated by
an appropriate amount. Let ǫ > 0 and denote by
v(x) = u(x − ǫen ).
Since u is strictly monotone in the vertical direction, v(x) < u(x) on CL+ (u). We
claim that
(2.2)
v(x) ≤ u(x) − c1 ǫ
on B9/10 (0) × {L − 1/2}
for ǫ < ǫn a dimensional constant, and a constant c1 > 0 depending only on L
and n. To prove (2.2), note first that from (2.1) it follows that u(x) ≥ bn for
all x ∈ B9/10 (0) × {L − 1/2}. Write x = (x′ , L − 1/2) and let tn be such that
u(x′ , tn ) = bn /2, then by monotonicity u(x′ , t) ≥ bn /2 for all t ≥ tn . Consider
the segment from (x′ , tn ) to (x′ , L − 1/2). It follows from the Lipschitz bound
(Lemma 2.2) that the distance from any point of the segment to the free boundary
is greater than a dimensional constant. Thus by Harnack’s inequality the values
of w(x) = (∂/∂xn )u(x) on this segment are comparable with a constant depending
only on n and L. Furthermore,
ˆ L−1/2
w(x′ , t)dt.
bn − bn /2 ≤ u(x) − u(x′ , tn ) =
tn
Therefore, the minimum of w on this segment is bounded below by a constant
c1 > 0, depending only on n and L. In particular,
ˆ L−1/2
u(x) − v(x) =
w(x′ , t)dt ≥ c1 ǫ.
L−1/2−ǫ
8
D. DE SILVA AND D. JERISON
Step 2: Construction of a family of supersolutions.
The hypothesis of Theorem 1.3 implies (by the construction of P. W. Jones [J])
that there is an NTA domain between any pair C(r1 , L − a1 ) and C(r2 , L − a2 )
for r1 < r2 ≤ 9/10 and a1 > a2 ≥ 1/2. Thus the boundary Harnack inequality,
Theorem 2.5, has the following corollary.
Corollary 2.8. Let u be as in Theorem 1.3 and let r1 < r2 ≤ 9/10 and a1 >
a2 ≥ 1/2. Then there is a constant A depending on L, the NTA constants of D,
r2 − r1 > 0, and a1 − a2 > 0 such that if h1 and h2 are positive harmonic functions
on C(r2 , L − a2 ) ∩ CL+ (u), vanishing on ∂D ∩ C(r2 , L − a2 ) then
h1 (x)/h2 (x) ≤ Ah1 (y)/h2 (y)
for every x and y in C(r1 , L − a1 ) ∩ CL+ (u).
In this step we start our analysis on the cylinder C(9/10, L − 1/2) which by
abuse of notation we denote by C1 . Then we restrict to smaller cylinders C2 , C3
with base B8/10 and B7/10 respectively, height M with L − 1 < M < L − 1/2 and
C3 ⊂⊂ C2 ⊂⊂ C1 .
Let w be the harmonic function in C1+ (u), satisfying the following boundary
conditions:
(2.3)
w = 0,
on F (u),
v < w ≤ u, on C1+ (u) ∩ ∂C1 ,
c1
c1
v + ǫ < w < u − ǫ, on B9/10 × {L − 1/2}.
(2.5)
4
4
Notice that (2.5) can be achieved because of the gap (2.2) between u and v.
Since v is subharmonic and u is harmonic in C1+ (u), the maximum principle implies
(2.4)
(2.6)
v<w<u
in C1+ (u).
Moreover, C1+ (w) = C1+ (u), and F (w) = F (u) ∩ C1 .
We claim next that in the smaller cylinder C 2 ,
(2.7)
|∇w|(x) ≤ C1 ,
x ∈ C 2.
Define d(x) = dist(x, F (u)). At points x ∈ C 2 ∩ C1+ (u) such that d(x) ≥ 1/10, this
follows from standard elliptic regularity and the fact that w is bounded. On the
other hand, at points that are close to F (u), we have that Bd(x) (x) ⊂ C1+ (u) and
from Lemma 2.2,
w(x) < u(x) ≤ Kd(x).
A standard argument using rescaling implies the bound (2.7).
Now, set h = u − w. Then h is a positive (see (2.6)) harmonic function on
C1+ (u) vanishing continuously on F (u). Let H be the harmonic function in the
cylinder B9/10 × (L − 1, L − 1/2), with boundary data c1 /2 on the top of the
cylinder and vanishing on the remaining part of the boundary. Then, in view
of (2.5), h ≥ ǫH. Thus, h(x1 ) ≥ c1 ǫ/4, at x1 = (L − 1/2 − δn )en for a small
dimensional constant δn > 0. Moreover, by the Lipschizt continuity of u we get
that h(x1 ) < (u − v)(x1 ) ≤ Kǫ. Using non-degeneracy and Lipschizt continuity of
u we also have that bn ≤ u(x1 ) ≤ 2LK. Thus, Corollary 2.8 gives
c2 ǫu ≤ h ≤ C2 ǫu on C2+ (u).
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
9
The upper bound on h implies,
(2.8)
w(x) ≥ (1 − C2 ǫ)u(x) on C2+ (u),
while the lower bound gives
(2.9)
w(x) ≤ (1 − c2 ǫ)u(x) on C2+ (u).
In particular, if F (u) is smooth around a point x0 ∈ C2 then |∇u|(x0 ) = 1, which
combined with (2.9) gives
(2.10)
|∇w|(x0 ) ≤ 1 − c2 ǫ.
According to Lemma 2.7 we then have
(2.11)
|∇w| ≤ 1 − c2 ǫ
ω-almost everywhere on F (u) ∩ C2 .
Next we use (2.11) to show that, by restricting on the smaller cylinder C3 , we
have
(2.12)
on C3+ (u).
|∇w| ≤ 1 − c2 ǫ + C3 u
Let h̃ be the largest harmonic function h̃ ≤ C1 in C2+ (u) such that
h̃ = 1 − c2 ǫ
on F (u) ∩ C2
with C1 the constant in (2.7). Since |∇w| is subharmonic, it satisfies (2.7)-(2.11)
we get
(2.13)
|∇w| ≤ h̃.
On the other hand, h̃ − (1 − c2 ǫ) is a positive harmonic function on C2+ (u), and
it is zero on F (u). Since by non-degeneracy u is bounded below by a dimensional
constant on the top of C3 , Corollary 2.8 gives
h̃ − (1 − c2 ǫ) ≤ C3 u
on C3+ (u).
Combining this inequality with (2.13) we obtain (2.12).
We now use (2.12) to construct a family of strict supersolutions. Define for t ≥ 0,
with
wt (x) = w(x) − tg(x),
x ∈ C1
g(x) = eAxn φ (|x′ |)
where A is a positive constant to be chosen later, and φ ≥ 0 is a smooth bump
function such that
1, if r < 1/2
φ(r) =
0, if r ≥ 7/10.
Moreover, we will choose φ such that φ(r) > 0 for r < 7/10
n−2 ′
φ (r) ≥ 0, if 6/10 ≤ r ≤ 7/10.
φ′′ (r) +
r
Indeed, let ψ(s) = e−2n/s for s > 0 and ψ(s) = 0 for s ≤ 0. Then for 0 ≤ s ≤ 1,
ψ ′′ (s) − 2nψ ′ (s) = [(2n + 4n2 )/s2 − (2n)2 /s]e−2n/s ≥ 0.
Because (n − 2)/r ≤ 2n for r ≥ 1/2, the function φ1 (r) = ψ(7/10 − r) satisfies the
differential inequality for φ above in the range r ≥ 1/2. Using a partition of unity,
φ1 can be modified without changing its values for r ≥ 6/10, to obtain a function
φ that is equal to 1 for r ≤ 1/2. Finally, using the inequalities for φ,
∆g = A2 eAxn φ (|x′ |) + eAxn ∆φ (|x′ |) ≥ 0
10
D. DE SILVA AND D. JERISON
as long as A is a sufficiently large dimensional constant.
Thus, wt is superharmonic on C1+ (wt ). Moreover, condition (2.12) together with
(2.8) imply that,
|∇wt | ≤ |∇w| + t|∇g| ≤ 1 − c2 ǫ + C4 w + t|∇g|,
on C3+ (u).
In particular, on F (wt ) ∩ C3 , t > 0, since w = tg we obtain
|∇wt | ≤ 1 − c2 ǫ + C4 tg + t|∇g|.
Therefore, for 0 < t ≤ c3 ǫ, with c3 small depending on c2 , C4 , and A, we deduce
that
c2
(2.14)
|∇wt | ≤ 1 − ǫ on F (wt ) ∩ C3 .
2
Step 3: Comparison.
Observe that because g vanishes on the “sides” we have that
(2.15)
wt = w > v
on (∂B9/10 × [−L, L]) ∩ C1+ (wt ),
and according to (2.5) we have that
(2.16)
wt > v
on B9/10 × {L − 1/2} for t ≤
c2 −A(L−1/2)
e
ǫ = c4 ǫ.
4
Let E = {t ∈ [0, c4 ǫ] : v ≤ wt in C1 }. We claim that E = [0, c4 ǫ]. Indeed, 0 ∈ E
and clearly E is closed. We need to show that E is open. Let t0 ∈ E, then since
wt is superharmonic in its positive phase and satisfies (2.15)-(2.16) we only need to
show that wt0 > v = 0 on F (v) ∩ C1 .
In the case t0 = 0, w0 = w > 0 on F (v) follows from the assumption that F (u)
is a graph in the vertical direction. In fact for all t, wt = w > 0 on F (v) ∩ (C1 \C3 )
because g is zero there. It remains to rule out the case, in which t0 > 0, and F (v)
touches F (wt0 ) in C3 , that is, where g(x0 ) 6= 0.
Suppose by contradiction x0 ∈ F (v) ∩ F (wt0 ) ∩ C3 , t0 > 0. If ∇wt0 (x0 ) 6= 0,
then by the implicit function theorem F (wt0 ) is smooth in a neighborhood of x0
and hence there exists an exterior tangent ball B at x0 for F (v). Therefore, for ν
the outward unit normal to B at x0 we have that
wt0 (x) ≥ v(x) = (x − x0 , ν)+ + o(|x − x0 |)
as x → x0 , contradicting (2.14).
On the other hand, if ∇wt0 (x0 ) = 0, then in a small neighborhood Br (x0 ) we
have that
wt0 (x) ≤ Cr2 .
However, according to the corkscrew condition, there exists a ball Bδr (y) ⊂ Br (x0 )∩
C1+ (v), for some small δ > 0. By the non-degeneracy of v we then obtain
sup v ≥ cr,
Br (x0 )
and again we reach a contradiction.
Thus c4 ǫ ∈ E, and
v ≤ wc4 ǫ on C 1 .
Hence, according to the definition of g,
{w ≤ c4 e−AL ǫ} ∩ {|x′ | < 1/2} ⊂ {v = 0} ∩ {|x′ | < 1/2}.
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
11
Moreover, by (2.7)
w ≤ C1 d(x).
Thus,
{d(x) ≤ c6 ǫ} ∩ {|x′ | < 1/2} ⊂ {v = 0} ∩ {|x′ | < 1/2}.
This implies the Lipschitz continuity of F (u) ∩ {|x′ | < 1/2} with bound depending
only on L, n and the NTA constants.
3. A priori gradient bound for minimal surfaces
In this section we present our proof of Theorem 1.2. Recall that Br denotes an
open (n − 1)-dimensional ball of radius r, while Br , denotes an open n-dimensional
ball of radius r.
Our proof is parallel to one in the free boundary setting above. One main
ingredient which will allow us to apply our deformation argument will be the (weak)
Harnack inequality for solutions to elliptic equations on minimal surfaces due to
Bombieri and Giusti [BG]. We recall its statement in the form in which we will use
it later in the proof.
Let ∆S denotes the Laplace-Beltrami operator on the surface S.
n−1
. There is a constant C(p) < ∞ and β > 0 depending
n−3
on dimension such that if S is an area minimizing hypersurface in BR = BR (x0 )
and x0 ∈ S and v is a positive supersolution to the Laplace-Beltrami operator,
∆S v ≤ 0, in BR ∩ S, then
1/p
p
(3.1)
≤ C(p) inf v
v dHn−1
Theorem 3.1. Let p <
Br ∩S
Br ∩S
for all r ≤ βR.
Corollary 3.2. Let S be an oriented surface of least area in B1 × R ⊂ Rn−1 × R.
Assume S1/2 := S ∩ (B1/2 × R) is connected, and S ⊂ B1 × [−M, M ]. Let v be a
positive supersolution to to the Laplace-Beltrami operator, ∆S v ≤ 0, in (B1 ×R)∩S,
such that
ˆ
(3.2)
vdHn−1 ≥ 1.
S1/2
Then
(3.3)
v≥c
with c > 0 depending only on n and M.
on
S1/2 ,
Proof. Let β (small) be the constant in Theorem 3.1. Decompose Rn into cubes of
√
ei ⊃ Qi
side-length β/(20 n). For each cube Qi that intersects S1/2 take a ball B
with center xi on S1/2 ∩ Qi and radius β/20. Clearly, the number N of balls Bei
that cover S1/2 depends only on n and M.
We say that Bei ∼ Bej if there exists a chain of balls Bek connecting Bei and Bej
such that consecutive balls intersect. This defines an equivalence relation. To each
equivalence class we can associate the open set which is the union of all the elements
in the class. Notice that open sets corresponding to distinct equivalence classes are
disjoint. Since S1/2 is connected, we conclude that all the balls belong to the same
equivalence class.
12
D. DE SILVA AND D. JERISON
If Be1 and Be2 intersect then they are both contained in Bβ/2 (x1 ). Hence applying
Theorem 3.1 we obtain
ˆ
ˆ
vdHn−1 ≤ C0
(3.4)
vdHn−1 ≤ C1 inf v ≤ C2
vdHn−1 .
e1 ∩S
B
Sβ/2 (x1 )
Bβ/2 ∩S
e2 ∩S
B
In the last inequality we used the well-known fact that
Hn−1 (S ∩ Bρ (x)) ≈ ρn−1
(3.5)
for all x ∈ S.
It also follows from (3.2) that at least one of the balls, say Be1 , satisfies
ˆ
vdHn−1 ≥ 1/N.
(3.6)
e1
S1/2 ∩B
Combining (3.4)-(3.6) with the fact that any two balls can be connected by a chain
of length at most N , we obtain the desired conclusion.
Proof of Theorem 1.2. In what follows, the constants c, ci , C, Ci depend only on
n and M . Denote by S the graph of φ over B1 . We present the proof in three steps.
Step 1: Separation on a set of substantial measure.
Let ǫ > 0 and set
Sǫ := {(x, φ(x) + ǫ) : x ∈ B1 }.
We will prove that there exists a smoothly bounded, closed set Ẽ ⊂ B1/2 of positive
measure independent of ǫ as ǫ → 0 such that
(3.7)
dist((x, φ(x)), Sǫ ) ≥ c0 ǫ
for all x ∈ Ẽ + Bδ
where δ > 0 depends on the (a priori) bound on the modulus of continuity of ∇φ.
Let η ∈ C0∞ (B1 ) be a smooth cut-off function such that η ≡ 1 on B1/2 . Then,
since φ satisfies (1.2) we have that
∇φ · ∇(η 2 φ)
p
dx = 0.
1 + |∇φ|2
ˆ
B1
Hence,
ˆ
|∇φ|2
η2 p
dx = − 2
1 + |∇φ|2
B1
2
ˆ
B1
ˆ
B1
Thus,
∇φ · ∇η
φη p
≤
1 + |∇φ|2
!1/2 ˆ
!1/2
η 2 |∇φ|2
φ2 |∇η|2
p
p
.
1 + |∇φ|2
1 + |∇φ|2
B1
|∇φ|2
η2 p
dx ≤ 4
1 + |∇φ|2
B1
ˆ
Since η ≡ 1 on B1/2 we then get
ˆ
(3.8)
B1/2
|∇η|2
φ2 p
dx ≤ CM 2
1 + |∇φ|2
B1
ˆ
|∇φ|dx ≤ C0 .
with C0 depending on M and n only. Hence, by Chebyshev’s inequality, (for C1 =
2C0 /|B1/2 |)
|{x ∈ B1/2 : |∇φ| < C1 }| ≥ |B1/2 |/2.
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
13
Since φ is smooth, there is a closed, smoothly bounded set
Ẽ ⊃ {x ∈ B1/2 : |∇φ| < C1 }
and δ > 0 sufficiently small depending on the modulus of continuity of ∇φ such
that
Ẽ + Bδ ⊂ {x ∈ B1/2 : |∇φ|2 ≤ C12 + 1}
This implies the desired claim (3.7), for small enough ǫ and δ, depending on the
smoothness of φ.
In what follows we denote by E = {(x, φ(x)), x ∈ Ẽ}. Clearly, Hn−1 (E) ≥
|B1/2 |/2.
Step 2: Construction of a family of subsolutions.
For the time being let S be any smooth surface. Denote by H(P, S) the mean
curvature of S at a point P ∈ S, (i.e. the trace of the second fundamental form of
S at P .) Assume that S is a smooth graph over B1 , i.e. S = {(x, φ(x)) : x ∈ B1 },
and let w be a C 2 non-negative function on S. Consider the surface St,ν := S + twν
obtained deforming S along the upward unit normal to S, that is
St,ν = {(x, φ(x)) + tw(x, φ(x))νx , x ∈ B1 },
with
(−∇φ(x), 1)
.
νx = p
1 + |∇φ(x)|2
Then, for t small enough, St is also a graph and one can compute (see for example
[K])
(3.9)
(3.10)
H(Pt , St,ν ) = H(P, S) + t(∆S w(P ) + |A|2S w(P )) + O(t2 ),
P := (x, φ(x)), Pt := (x, φ(x)) + tw(x, φ(x))νx ,
where |A|S is the norm of the second fundamental form of S. (The O(t2 ) term
depends at most on the third derivatives of φ and on the second derivatives of w.)
Applying formula (3.9) to our minimal surface S we find that
(3.11)
H(Pt , St,ν ) = t(∆S w(P ) + |A|2S w(P )) + O(t2 ).
In order to run a continuity argument (as in the proof of Theorem 1.3), we
wish to use formula (3.11) to produce a family of surfaces St,ν which are strict
subsolutions to the minimal surface equation i.e. H(·, St,ν ) > 0 at least outside
Et,ν := E + twν, with E the set from the previous step. Towards this aim we prove
the following claim.
Claim. There exists a function w defined on S such that
∆S w + |A|2S w > 0
on S \ E,
w(x, φ(x)) = 1 on Ẽ,
w(x, φ(x)) = 0 on ∂B1 .
Moreover
(3.12)
w(x, φ(x)) ≥ c0 > 0
on B1/2 ,
with c0 depending only on n, M and w ∈ C 2 (S \ E) with C 2 bounds depending on
S and E.
14
D. DE SILVA AND D. JERISON
Proof of the claim. Let w1 be the solution to the following boundary value problem,
∆S w1 = 0 on S \ E,
w1 (x, φ(x)) = 1
on ∂ Ẽ,
w1 (x, φ(x)) = 0
on ∂B1 .
Note that the solution exists and is smooth in its domain of definition because Ẽ
is smoothly bounded. Extend w1 = 1 on Ẽ. Then ∆S w1 ≤ 0 on S. Moreover,
according to Step 1, we have that (using the notation of Corollary 3.2)
ˆ
ˆ
w1 dHn−1 ≥
w1 dHn−1 = Hn−1 (E) ≥ |B1/2 |/2.
S1/2
E
Hence we can apply Corollary 3.2 to conclude that
(3.13)
w1 ≥ c
on S1/2 = S ∩ (B1/2 × R).
Now, let w0 be the solution to the following problem:
∆S w0 = 1 on S \ E,
w0 (x, φ(x)) = 0
on Ẽ,
w0 (x, φ(x)) = 0
on ∂B1 .
and set w = w1 + δ1 w0 . Clearly, |∇w0 | is bounded (by a constant depending on
S and E). Applying Hopf’s lemma to w1 on (∂B1 × R) ∩ S , we obtain that, for
δ1 sufficiently small, w > 0 in a neighborhood of (∂B1 × R) ∩ S and hence (for a
possibly smaller δ1 ) w > 0 on S. Moreover, in view of (3.13), we can choose δ1 so
that w satisfies (3.12). Thus, w has all the required properties.
In view of the claim, according to formula (3.11), if t is sufficiently small, 0 <
t ≤ ǫ0 then
H(·, St,ν ) > 0, on St,ν \ Et,ν .
Step 3: Comparison.
We show that for 0 ≤ t ≤ c0 ǫ ≤ ǫ0 , the surface St,ν is below the surface Sǫ .
Indeed, this is true at t = 0. The first touching point cannot occur at some x ∈ ∂B1 ,
as our deformation leaves the ∂B1 fixed. Moreover, for t small enough, no touching
can occur on Et,ν in view of (3.7) in Step 1. Finally St,ν is a strict subsolution on
St,ν \ Et,ν , hence no touching can occur there either. Since w satisfies (3.12), we
can then conclude that for all sufficiently small ǫ, (recall Sc0 ǫ,ν = S + c0 ǫwν)
dist((x, φ(x)), Sǫ ) ≥ dist((x, φ(x)), Sc0 ǫ,ν ) ≥ c1 ǫ
on B1/2
as desired. Note that although the size of ǫ0 depends on the a priori bound on ∇φ,
the constants c0 > 0 and c1 > 0 do not.
4. Final Remarks
The analogy between the two gradient bound proofs presented here goes farther.
Not only does each proof depend crucially on a scale-invariant Harnack inequality
for the second variation operator of the associated functional, but also the proofs
of these two Harnack estimates follow a roughly parallel course.
The key ingredient of our proof of the gradient bound for minimal surface graphs
is the Harnack inequality for the Laplace-Beltrami operator on the surface. This
A GRADIENT BOUND FOR FREE BOUNDARY GRAPHS
15
Harnack inequality permits us to convert a gradient bound on average (separation
on a set of substantial measure) to a gradient bound everywhere (separation everywhere). The way this Harnack inequality is proved by Bombieri and Giusti is as
follows. A monotonicity formula yields (via a limiting cone argument) a measuretheoretic form of connectivity. This, in turn, implies another scale-invariant form
of connectivity, an isoperimetric, or Poincaré-type, inequality. One then deduces a
Harnack inequality for the Laplace-Beltrami operator on the minimal surface by a
Moser-type argument.
In the free boundary case, a monotonicity formula due to Alt, Caffarelli and
Friedman yields (by arguments of [ACS] and [D1]) the NTA property, a scaleinvariant form of connectivity. A theorem of [JK] says that the NTA property
implies a boundary Harnack inequality. The boundary Harnack inequality is used
to show that separation of level surfaces of the solution function u at distances
far from the free boundary implies a similar separation all the way up to the free
boundary.
The parallel between these two Harnack inequalities leads to the hope that there
is a Harnack estimate for the second variation operator associated to minimizers of
functionals of the form
ˆ
|∇v|2 + F (v)
for wider classes of functions F .
References
[AC]
Alt H.W., Caffarelli L.A., Existence and regularity for a minimum problem with free boundary, J. Reine Angew. Math 325 (1981),105–144.
[ACS] Aguilera, N. E., Caffarelli, L. A., and Spruck, J., An optimizatiion problem in heat conduction, Ann. Scuola Norm. Sup. Pisa Cl. Sc (4) 14 (1988) 355–387.
[BDG] Bombieri E., De Giorgi E., Giusti E., Minimal cones and the Bernstein problem, Inv.
Math. 7 (1969), 243–268.
[BDM] Bombieri E., De Giorgi E., Miranda M., Una maggiorazione a priori relativa alle ipersuperfici minimali non parametriche, (Italian) Arch. Rational Mech. Anal. 32 (1969) 255–267.
[BG] Bombieri E., Giusti E., Harnack’s inequality for elliptic differential equations on minimal
surfaces, Inventiones Math. 15 (1972) 24–46.
[C1] Caffarelli L.A., A Harnack inequality approach to the regularity of free boundaries. Part I:
Lipschitz free boundaries are C 1,α , Rev. Mat. Iberoamericana 3 (1987), no.2, 139–162.
[C2] Caffarelli L.A., A Harnack inequality approach to the regularity of free boundaries. Part
II: Flat free boundaries are Lipschitz, Comm. Pure Appl. Math. 42 (1989), no. 1, 55–78.
[C3] Caffarelli L.A., A Harnack inequality approach to the regularity of free boundaries. Part
III: Existence theory, compactness, and dependence on X, Ann. Scuola Norm. Sup. Pisa
Cl. Sci. (4) 15 (1988), no. 4, 583–602.
[CJK] Caffarelli L.A., Jerison D., Kenig C.E., Global energy minimizers for free boundary problems and full regularity in three dimension, Noncompact problems at the intersection of
geometry, analysis, and topology, 83–97, Contemp. Math., 350, Amer. Math. Soc., Providence, RI, 2004.
[CS] Caffarelli L.A., Salsa S. A geometric approach to free boundary problems, Graduate Studies
in Mathematics, 68. Amer. Math. Soc., Providence, RI, 2005.
[D1] De Silva D., Existence and regularity of monotone solutions to free boundary problems,
Amer. J. Math. 131 (2009), no.2, 351–378.
[D2] De Silva D., Bernstein-type techniques for 2D free boundary graphs, Math. Z. 260 (2008),
no. 1, 47–60.
[DJ] De Silva D., Jerison D., A singular energy minimizing free boundary to appear in J. Reine
Angew. Math.
[JK] Jerison D., Kenig C.E., Boundary behavior of harmonic functions in nontangentially Accessible Domains, Adv. in Math. 46 (1982), no. 1, 80–147.
16
[J]
[KN]
[K]
D. DE SILVA AND D. JERISON
Jones P.W., A geometric localization theorem,, Adv. in Math. 46 (1982), 71–79.
Kinderlehrer D., Nirenberg L., Analyticity at the boundary of solutions of nonlinear secondorder parabolic equations, Comm. Pure Appl. Math. 31 (1978), no. 3, 283–338.
Korevaar, N., An easy proof of the interior gradient bound for solutions to the prescribed
mean curvature equation,, Proc. Symposia Pure Math. 45 (1986), Part 2, 81–89.
Department of Mathematics, Barnard College, Columbia University, New York, NY
10027
E-mail address: desilva@math.columbia.edu
Department of Mathematics, Massachusetts Institute of Technology, Cambridge,
MA
E-mail address: jerison@math.mit.edu